What potential for the NV31 and NV34?

[Tahir2]

I hope the R350 doesnt magically disappear off the map but looking at the current numbers does ATI need to get it out desparately? IMO no.

RV350 will probably be as interesting as the Radeon 9000 Pro and we have a repeat of history.

If ATI is serious though then R350 should be able to beat the NV30 by a significant margin and be significantly faster (I say 20% at least) than a R300 for it to be worth bringing out.

 

[Grall]

Chal:

"My first guess would be that this is a memory granularity issue. The Radeon 9700 Pro uses 4 128-bit memory controllers, while the GeForce FX has 4 64-bit controllers."

I wouldn't bet my life on that. Doesn't DDR2 memory have a burst rate of 4 compared to standard DDRs 2? Ie, 4x32 bit memory chips transfers a total of 16 32-bit words per clock, or 512 bits; the exact same as the R300...

I suppose Codecreatures either is very friendly to Nvidia's design in general or is plain fillrate-limited, so that the NV30s high core clock makes a difference. It's difficult to draw any real conclusions from a synthetic benchmark after all...


*G*

 

[Clashman]

Isn't it more like GFFX's 4x AA can usually outperform the 9700's 6x AA? I don't think any of the reviews have actually compared 4 to 4. And the ones that have compared GFFX's 4 to ATI's 6 have said that the ATI method wins hands down in IQ.

Huh

Reviews I have read difinately compared 4x to 4x.
ATI's 6x will of course wins hands down to Nv's 4x.

I am not sure what you are getting at??? :?

My apologies. I don't know what I was on.

 

[Clashman]

Also, I don't think this (relatively speaking) poor showing by nVidia will affect whether or not R350 comes out. There are internal reasons for bringing it to market regardless of how well R300 stacked up to GFFX. Namely profits. A retooled version of the R300 that's cheaper to produce, consumes less power and produces less noise noise, and performs well enough to allow them to charge $350 for it will be a much higher margin product than the current R300, and so they have every reason to bring it out ASAP.

 

[jb]

I don't think so. Looking at Tomshardware and Hardocp, the GeForce FX wins until the resolution is cranked up too high (where both are below 60 fps...). This, to me, looks like a clear driver issue.

At those settings its clearing hurting for bandwidth. I think you can get some speed bumb with drivers but not much in these cases. When you dont have the bandwidth you have to work that much harder for every gain. Are you also forgetting the massive fill rate advatage that the FX has?

but, at the same time, right now the FX is quite a bit better than the Radeon 9700 for older games with alpha test.

Only for those games that are D3D

 

[antlers]

I don't know if the R350 will be cheaper to produce than the R300. The chip might be (depending on what they've done to it) but the memory will certainly be more expensive.

 

[MuFu]

A 400MHz-450MHz NV31 was part of nVidia's fall lineup. They do not have one at this time.

Subsequentally, it looks like there will be no card capable of competing with 9500 Pro and 9700 in terms of price/performance. An NV31 at 300MHz-350MHz is simply not fast enough. I doubt they can respin it again given time/cost constraints.

Final R350 Si coming back soon, btw...

MuFu.

Some answers (kind of)...

NV36 helps bridge the gap between NV31 and NV30. It was drafted in only a few months ago so is most likely a mildly toned-down NV30. There may be a mobile derivative in development also.

The NV30 PCB initially had 8 layers. Then two more were added to help SSTL_18 integrity issue. Then two more... make of that what you will. Certainly seems to support the idea that they "overclocked" the card more and more as development progressed.

NV35 had 10 initially - now it is most likely 12 or more (omfg).

MuFu.

 

[Mulciber]

Chal:

"My first guess would be that this is a memory granularity issue. The Radeon 9700 Pro uses 4 128-bit memory controllers, while the GeForce FX has 4 64-bit controllers."

I wouldn't bet my life on that. Doesn't DDR2 memory have a burst rate of 4 compared to standard DDRs 2? Ie, 4x32 bit memory chips transfers a total of 16 32-bit words per clock, or 512 bits; the exact same as the R300...

I suppose Codecreatures either is very friendly to Nvidia's design in general or is plain fillrate-limited, so that the NV30s high core clock makes a difference. It's difficult to draw any real conclusions from a synthetic benchmark after all...


*G*

a comon misconception about DDRII. The burst rate they are refering to is between the bit array and the memory bus. clock for clock DDR and DDRII produce the same ammount of bandwidth per pin. The difference in latency between the two IMHO should be negligable considering that the NV30s internal times are tuned for DDRII and that most texture reads are hardly random, unlike the random strains that a CPU will put on your system memory.

 

[Mulciber]

A 400MHz-450MHz NV31 was part of nVidia's fall lineup. They do not have one at this time.

Subsequentally, it looks like there will be no card capable of competing with 9500 Pro and 9700 in terms of price/performance. An NV31 at 300MHz-350MHz is simply not fast enough. I doubt they can respin it again given time/cost constraints.

Final R350 Si coming back soon, btw...

MuFu.

Some answers (kind of)...

NV36 helps bridge the gap between NV31 and NV30. It was drafted in only a few months ago so is most likely a mildly toned-down NV30. There may be a mobile derivative in development also.

The NV30 PCB initially had 8 layers. Then two more were added to help SSTL_18 integrity issue. Then two more... make of that what you will. Certainly seems to support the idea that they "overclocked" the card more and more as development progressed.

NV35 had 10 initially - now it is most likely 12 or more (omfg).

MuFu.

Ok now I'm confused lol.. So the NV36 is going to replace a scrapped nv34, or is the nv36 a fall 03 product?

And Im still hoping that nv35 has a 256bit bus

 

[Mr.huang]

http://www.hardware.fr/articles/453/page9.html

wow R9500 Pro

 

[Dave Baumann]

NV36 helps bridge the gap between NV31 and NV30. It was drafted in only a few months ago so is most likely a mildly toned-down NV30. There may be a mobile derivative in development also.

That sounds like they need something to compete with 9500 PRO. I.e. it could be a lower clock, cheaper variant of NV30 with 8 pipes or so. However, that also sounds quite concerning if it’s the case. A reactionary position is not necessarily a good one.

NV35 had 10 initially - now it is most likely 12 or more (omfg).

If it’s a 256bit bus and similar / higher speeds than NV30 then at this point I’d have difficulty seeing it being less.

 

[Mulciber]

NV36 helps bridge the gap between NV31 and NV30. It was drafted in only a few months ago so is most likely a mildly toned-down NV30. There may be a mobile derivative in development also.

That sounds like they need something to compete with 9500 PRO. I.e. it could be a lower clock, cheaper variant of NV30 with 8 pipes or so. However, that also sounds quite concerning if it’s the case. A reactionary position is not necessarily a good one.

NV35 had 10 initially - now it is most likely 12 or more (omfg).

If it’s a 256bit bus and similar / higher speeds than NV30 then at this point I’d have difficulty seeing it being less.

Seems to me that the only thing they could really do to make it cheeper but still perform at 9500pro levels would be to make it an 8 layer pcb and put a very inexpensive HSF of it, at around 300 or 350mhz. And even then it would still not be able to compete with ATIs "lesser" quality anisotropic performance. :?

 

[Dave Baumann]

Seems to me that the only thing they could really do to make it cheeper but still perform at 9500pro levels would be to make it an 8 layer pcb and put a very inexpensive HSF of it, at around 300 or 350mhz.

Well, going by the clock speeds of Quadro FX, given that even the 300/300 version is still using DDR-II I'd guess the NV30/NV30GL memory interface is DDR-II only. Moving back to a DDR-I interface would make things cheaper using 300MHz DDR-I RAM, and the low core/mem speeds would almost certianly facilitate few layers. I should imagine theres other stuff they can remove at a silicon level as well, if needs be.

 

[Snyder]

http://www.hardware.fr/articles/453/page9.html

wow R9500 Pro

Thanks for the link, this is in fact a quite interesting benchmark (although with my limited ability of reading French I coulnd't discover which AF method was used for both cards), because they downclocked a GFFX to the same values as a R9500 Pro, so what we have here is an exact clock-for-clock comparison of both chips. (Both 275/270,8x1,128Bit).

If the benches are somewhat reliable, the FSAA/AF performance of the R300 is quite impressive.

 

[MuFu]

Ok now I'm confused lol.. So the NV36 is going to replace a scrapped nv34, or is the nv36 a fall 03 product?

I don't think it is being scrapped, no. I guess we'll see an NV31 clocked at about 300MHz or so released in late March along with NV34 (the most severely retarded of the NV30 derivatives). It'll be competitive with the 9500NP, leaving a big gap between it and the 5800NU which will be addressed by NV36 at a later date. I wouldn't be suprised if NV30 is then made immediately obsolete by NV35. A 4-ASIC lineup is pretty ridiculous as it is, not to mention the stockpiles of NV25/NV28s that will continue to sell for a while yet.

Speaking of memory interfaces - a new DDR-I controller/interface was developed for NV31 so maybe NV36 marries that with the NV30 pipeline (or something very similar).

It's quite possible that the NV35 board will have 12 layers (and no more). The extra grounding layers in the NV30 PCB were probably added in a "better safe than sorry" manner, given time constraints. PCBs are reletively inexpensive and new ones can be turned over pretty quickly (I believe). I'm sure it is possible to develop a board that has a 256-bit bus and despite another clockspeed hike to 600-700MHz (further increasing capacitive/inductive coupling probelms), still has "only" 12 layers a la NV30.

MuFu.

 

[Joe DeFuria]

Well, with 4x FSAA, the GeForce FX is usually still able to perform better than the Radeon 9700 Pro, so the NV31/NV34 shouldn't have much trouble against the 9500 series. Don't forget that the Radeon 9500 doesn't have HyperZ III enabled.

Well, Chalnoth, in light of the recent "apples to apples" benchmarks of an 8 pipe NV3X core vs. the 9500 Pro....care to revise your statement about the NV31/34 not having "much trouble" against the 9500 series?

 

[MuFu]

Heh.

Well, with 4x FSAA, the GeForce FX is usually still able to perform better than the Radeon 9700 Pro, so the NV31/NV34 shouldn't have much trouble against the 9500 series. Don't forget that the Radeon 9500 doesn't have HyperZ III enabled.

Shouldn't have trouble?!!! NV34 is not competitive with the Radeon 9500/9500NP at all - I expect it to be only slightly faster than a Radeon 9000 Pro and "DX9 compliant". The NV31 may be *if* they can get clockspeeds above 350MHz, which looks near-on impossible right now. At 300Mhz it should be about on a par with the 9500NP - maybe slightly slower overall. Unfortunately you can't software mod it to a 5800 Ultra specs either...

MuFu.

 

[martrox]

http://www.hardware.fr/articles/453/page9.html

wow R9500 Pro

OUCH! nVidia's in major trouble......... This is really sad., and in no way helps anyone. It may be a year till nVidia's competitive again, if ever.

 

[Dave Baumann]

Thats with all the Intellisample techinques as well. Some of these may be dropped for NV31.

 

[CMKRNL]

Yes, NV34 is definitely coming out. That's their low cost, mass consumer part with both desktop and mobile variants. NV36 as far as I know is an NV35 derivative and is still some time away.

Most (if not all) of the Z and color compression HW is still present in NV31. NV34 has a lot of these features removed for cost reduction though. From what I know of the architectures however, I expect RV350 to outperform NV31 in most situations, especially with AA+AF dialed up.